I am wondering if anyone has a good explanation why some softwares, such as AFQ, often have trouble reconstructing the right AF? Is it more physiological (right AF not being as developed as left due to left-sided language dominance) or methodological (e.g. crossing fibers or other factors lower FA or CSD power; endpoint ROIs not well defined, etc)? I’m really curious, because I have seen this phenomenon mentioned in papers, but there has not been to my knowledge a publication revolving around this issue.
I believe (as you note) that it has been established that there are asymmetries in the Arcuate Fasciculus, so that may be part of the issue.
But to your broader question, have you tried methods other than AFQ? To my understanding, AFQ uses a dual way-point ROI based approach (or at least it used to) using the Mori atlas as a guide. As such, if the warp goes bad with that one of the ROIs for the Arcuate may be misplaced. I have also noticed AFQ failing to segment tracts like this (maybe also the IFOF?), but I didn’t investigate any regularities with it.
One good sanity check (that kind of gets at a presumption in your question) is to determine whether valid candidate streamlines for the Arcuate Fasciculus exist in your input tractogram. For this I recommend using what I call a “categorical segmentation” (in essence an exhaustive segmentation that divides in to categories, not unlike most connectomic approaches). There’s a brainlife.io app example herevin which the Arcuate candidates would be found in the “fronto-temoporal” category (in that the relevant streamlines connect the temporal and frontal lobes). By using something like this, you can see whether it’s your tractography in the first place, or whether it is the segmentation failing to find the structure.
Alternatively, you can try other segmentation methods (like mine, which can be found here) to see if its the input tractogram or the segmentation method.
I have tried other methods with varying success. Sometimes DIPY recobundles does not resolve right AF. TractSeg tends to get it consistently, but given it is based on a deep learning generative model, I am not surprised. Might investigate some other algorithms in the future.